/external/tensorflow/tensorflow/compiler/xla/service/ |
D | hlo_execution_profile.cc | 52 const HloCostAnalysis& cost_analysis, in CreateHloProfilePrinterData() argument 103 instruction_info->set_flop_count(cost_analysis.flop_count(*hlo)); in CreateHloProfilePrinterData() 105 cost_analysis.transcendental_count(*hlo)); in CreateHloProfilePrinterData() 106 instruction_info->set_bytes_accessed(cost_analysis.bytes_accessed(*hlo)); in CreateHloProfilePrinterData() 108 cost_analysis.optimal_seconds(*hlo)); in CreateHloProfilePrinterData()
|
D | hlo_execution_profile_test.cc | 53 HloCostAnalysis cost_analysis(shape_size_function); in TEST_F() local 56 CreateHloProfilePrinterData(profile_index_map, cost_analysis, in TEST_F()
|
D | memory_space_assignment.h | 91 const HloCostAnalysis& cost_analysis, 95 const HloCostAnalysis& cost_analysis() const { return cost_analysis_; } in cost_analysis() function 156 const HloCostAnalysis& cost_analysis, in MemorySpaceAssignmentCostAnalysis() argument 162 : cost_analysis_(cost_analysis), in MemorySpaceAssignmentCostAnalysis() 309 const MemorySpaceAssignmentCostAnalysis& cost_analysis, 802 const MemorySpaceAssignmentCostAnalysis& cost_analysis,
|
D | hlo_execution_profile.h | 100 const HloCostAnalysis& cost_analysis, const string& entry_computation_name);
|
D | memory_space_assignment.cc | 135 const HloCostAnalysis& cost_analysis, in Create() argument 144 cost_analysis, async_copy_bandwidth_bytes_per_second, in Create() 374 const MemorySpaceAssignmentCostAnalysis& cost_analysis, in CostAnalysisPrefetchIntervalPicker() argument 379 cost_analysis.hlo_live_range().instruction_schedule().size(), 0), in CostAnalysisPrefetchIntervalPicker() 381 cost_analysis.hlo_live_range().instruction_schedule().size(), 0), in CostAnalysisPrefetchIntervalPicker() 382 cost_analysis_(cost_analysis), in CostAnalysisPrefetchIntervalPicker() 2630 const MemorySpaceAssignmentCostAnalysis& cost_analysis, in GetMemoryBoundednessBufferIntervalCompare() argument 2632 return [&cost_analysis, cache](const BufferInterval& x, in GetMemoryBoundednessBufferIntervalCompare() 2634 float x_memory_boundedness = cost_analysis.GetMemoryBoundedness(x, cache); in GetMemoryBoundednessBufferIntervalCompare() 2635 float y_memory_boundedness = cost_analysis.GetMemoryBoundedness(y, cache); in GetMemoryBoundednessBufferIntervalCompare()
|
D | memory_space_assignment_test.cc | 56 auto cost_analysis = MemorySpaceAssignmentCostAnalysis::Create( in AssignMemorySpaceUsingCostAnalysis() local 62 *cost_analysis, /*min_async_copy_to_overlap_ratio=*/0.8, in AssignMemorySpaceUsingCostAnalysis() 68 *cost_analysis, &cache_), in AssignMemorySpaceUsingCostAnalysis() 321 Create(const HloCostAnalysis& cost_analysis, const HloModule& module) { in Create() argument 328 cost_analysis, /*async_copy_bandwidth_bytes_per_second=*/1, in Create() 382 const HloCostAnalysis& cost_analysis, in FakeMemorySpaceAssignmentCostAnalysis() argument 389 cost_analysis, async_copy_bandwidth_bytes_per_second, in FakeMemorySpaceAssignmentCostAnalysis() 4111 TF_ASSERT_OK_AND_ASSIGN(auto cost_analysis, in TEST_P() 4114 cost_analysis->SetOverrideForGetAsyncCopyElapsed([](const Shape& shape) { in TEST_P() 4119 *cost_analysis, in TEST_P() [all …]
|
/external/tensorflow/tensorflow/compiler/xla/service/cpu/ |
D | parallel_task_assignment.cc | 56 std::unique_ptr<HloCostAnalysis> cost_analysis) in DefaultCostModel() argument 59 cost_analysis_(std::move(cost_analysis)) {} in DefaultCostModel() 116 auto cost_analysis = absl::make_unique<HloCostAnalysis>(shape_size); in ParallelTaskAssignment() local 118 Status status = computation->root_instruction()->Accept(cost_analysis.get()); in ParallelTaskAssignment() 122 std::move(cost_analysis))); in ParallelTaskAssignment()
|
D | cpu_compiler.cc | 562 HloCostAnalysis cost_analysis(shape_size_bytes); in CreateHloProfilingArtifacts() local 563 TF_RETURN_IF_ERROR(entry_computation.Accept(&cost_analysis)); in CreateHloProfilingArtifacts() 565 **hlo_profile_index_map, cost_analysis, entry_computation.name()); in CreateHloProfilingArtifacts()
|
/external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
D | gpu_compiler.cc | 884 HloCostAnalysis cost_analysis(ShapeSizeBytesFunction()); in RunBackend() local 885 cost_analysis.set_bytes_per_second( in RunBackend() 887 TF_RETURN_IF_ERROR(module->entry_computation()->Accept(&cost_analysis)); in RunBackend() 890 cost_analysis.bytes_accessed()); in RunBackend() 894 CreateHloProfilePrinterData(*profile_index_map, cost_analysis, in RunBackend()
|